MLOps.community

Getting Humans Out of the Way: How to Work with Teams of Agents

·51 min·5 clips
How can you make AI agents create their own feature walkthroughs with screenshots to avoid manual QA?
It starts with screenshots. The first move is to find the actual error and zoom in on it, because the less visual noise an agent has to process, the easier it is to tell whether a change matters or just looks busy. That leads to reporting. The guest likes tools that automatically summarize the important parts of a run and surface what was hard, because an agent can usually name its own friction if the prompt leaves room for it. Erica comes up next. She downloaded skills built from the Coding Agents conference live stream transcripts, fed them into Cloud, and the feedback lands in a very practical place: they seem useful, and at the very least they are not making anything worse. The host keeps pushing on the visual regression idea. The guest answers by treating the skill as part of a larger discipline, where agents are not just told what to do but are also taught how to report up when the task gets messy. Then the conversation widens. Brumi makes it quick to spin out agents, and the guest describes giving five of them different angles on the same task, then reading their work trees, their verification results, and the design choices they produced. Picking one is the easy part. The harder part is learning how to compare the runs without pretending they were all the same. Merge conflicts take the next turn. The guest says humans experience them as a pain point, but that agents can be told to pull the latest from main, inspect the pull requests they are merging with, understand what each one is trying to do, and decide how to resolve the conflict. That is the bridge. It turns merging from a purely manual burden into a workflow where the model can do the reading before a person has to do the arbitration. The close shifts to provider choice. Even if Claude's own app were amazing, the guest says he would still want the option to move to Codex, keep Gemini in view, and avoid being stuck if the landscape changes. Google stays in the frame. The point is simple enough: the company looks well run, the people sound smart, and benchmarks alone do not settle how something feels in day-to-day use.

As heard by us

A practical look at agents, diffs, parallel work trees, and provider flexibility.

The episode stays practical, centering on ways to narrow an agent's attention through pixel diffs, zooming in on the actual error, and reports that surface the most important points.

Read the full review in PlayNext →

Why you'd press play

For when you want your agents to spot the actual error before you do.

Read the full recommendation in PlayNext →
Listen to the show on